A hybrid neural-fuzzy-neural framework for speech recognition

A hybrid neural-fuzzy-neural framework based on an understanding of human cognition processes is proposed. The basic framework consists of a self-organization feature map as a feature projector and a set of multilayer perceptron sub-nets as processors. This framework can cope with the ambiguous information comprising the outputs of the projector as well as a fuzzy decision making procedure. A generalized structure is presented. Initial experiments have shown that such a structure has strong ability to deal with the problem of overlapping speech features